Repository files navigation

Research into controllers for 2d and 3d Active Ragdolls (using MujocoUnity+ml_agents)

• Joe Booth (SohoJoe)

• Type: Discrete 2D
• Build (MacOS):v0.004Playable
• Actions: No-op, Forward, Backwards, Jump, Jump+Forward, Jump+Backwards
• Controls: Left arrow, Right arrow, Spacebar
• Mujoco Model: DeepMindHopper
• Hypostheis: Use discreate + random trainer to create human controller.
• Outcome:
  • SUCCESS - contriol feels responsive
  • ... Has emerging functionality - i.e. tap left for small step, swap direction in air

• Type: Continuous 2D
• Build (MacOS):v0.003
• Actions: Forward / Backwards
• Mujoco Model: DeepMindHopper
• Hypostheis: Use an adversarial hierarchical trained agent as the controller which gets the inverse reward of the locomation agent on a slower time step. The idea is that it will push the locomoation agent to focus on its weakest areas.
• Outcome:
  • FAIL - training is too heavily influenced by the number of steps the controller agent takes between decisions;
  • ... it maybe better to train a seperate agent on hyper-parms (i.e. meta learning)
  • ... having read more about these approaches (MAML, RL2, etc) it would be better to move to a Discreate conrtroller as ml-agents LSTM does not work well with Continuous actions.

• Type: Continuous 2D
• Build (MacOS, Windows):v0.002
• Actions: Forward / Backwards
• Input: Unity Axis input (left/right or a/d or joystick)
• Mujoco Model: DeepMindHopper
• Hypostheis: Use ml-agent (player + Heuristic) and unity input to create a player contoller .
• Outcome:
  • Works - makes a simple 2d proof of concept of 2d active ragdoll using RL.
  • I needed to use Discrete target velocities for stable training (-1,1,0) however it is continuous under player control.
  • Training was sensitive to the number of steps between chaning the input (see counter logic below)
• Notes:
  • Controller002InputBrain
    • BrainType: Player = input is from player
    • BrainType: Heuristic = input is from Controller002InputBrain.cs
  • Controller002InputDecision.cs -
    • Waits until counter is 0 then:
    • Takes random action and outputs AxisX as -1, 0, 1 (used as target velocity), then:
    • Set counter random 40-240
  • Controller002InputAgent.cs -

• Type: Discrete 2D
• Build (MacOS):v0.001
• Actions: Forward / Backwards
• Mujoco Model: DeepMindHopper
• Hypostheis: It should be simple to train a backwards / forwards by giving the agent a +1 / 1 velocity target which feeds the reward function.
• Outcome: It worked - but was harder than expected to find good hyper-parms. Next, wrap this in a proper ml-agent controller and allow user input.